In deep learning, we often assume that a model's performance is bound by the dimensions of its training data. If you train a robot to navigate a $10 \times 10$ room, you expect it to fail in a $100 \times 100$ warehouse. However, a growing body of research suggests that if a model learns the local logic of a task rather than memorizing global coordinates, it can scale to massive environments without a single extra step of training.
This property, often called spatial extensibility or zero-shot scaling, allows models to be trained on small, computationally cheap "windows" and deployed on global-scale problems that would otherwise be memory-prohibitive.
The Locality Principle
The core intuition behind this research is that many complex systems—from fluid dynamics to urban traffic—are governed by local interactions. A water molecule doesn't need to know the size of the ocean to know how to react to its immediate neighbors.
By using architectures with strong inductive biases like translation-equivariance, researchers can capture these local rules. Ivan Oseledets and Ivan Sosnovik demonstrated this in Neural networks for topology optimization, where fully convolutional layers trained on low-resolution grids successfully designed structures at much higher resolutions. Because the convolutional filters act as local feature detectors, they remain valid regardless of how many times they are tiled across a larger grid.
This isn't limited to static designs. In physics simulations, Omer Rochman Sharabi and Gilles Louppe showed in A Neural Material Point Method for Particle-based Simulations that fully convolutional emulators can generalize to larger or differently shaped spatial domains. The model learns the "physics" of the patch, and the patch can be any size.
Learning Algorithms Instead of Patterns
The most striking aspect of this direction is that it applies to planning and reasoning. If a model learns the algorithm for solving a problem, the size of the problem becomes a matter of computational budget rather than model capacity.
Bernardo Esteves and colleagues introduced NeuralSolver, an architecture that learns general algorithms on small-scale problems and executes them on scales orders of magnitude larger. Similarly, Peter Karkus and David Hsu developed QMDP-Net, which embeds a differentiable planner into a neural network. By increasing the number of planning iterations at inference time, a model trained on small simulated grids can navigate massive real-world maps zero-shot.
This "algorithmic" scaling is also seen in multi-agent systems. Hanmo Chen and Stone Tao found that agents trained with convolutional policies in Emergent collective intelligence from massive-agent cooperation and competition could transfer their coordination strategies from small maps to large ones. The agents aren't learning "where to go" in a global sense; they are learning "how to interact" with their neighbors.
Theoretical Foundations of Transferability
Why does this work so reliably for some tasks and not others? The answer lies in the relationship between shift-equivariance and stationarity.
In Transferability of Convolutional Neural Networks in Stationary Learning Tasks, Charilaos Kanatsoulis and Alejandro Ribeiro provide a mathematical framework for this phenomenon. They argue that if the underlying signal statistics are stationary—meaning the "rules of the game" are the same everywhere—then a CNN trained on a small window can be deployed on an arbitrarily large spatial domain with bounded performance loss.
To push this further, some researchers are moving beyond simple translations to more complex symmetries. Maksim Zhdanov and Erik J Bekkers developed AdS-GNN, which uses conformal equivariance to handle scaling and non-affine transformations. This allows models to "discover" physical constants and maintain performance across varying object sizes and perspectives without needing scale-based data augmentation.
Bridging the Local-Global Gap
While local rules are powerful, some problems require global context. The challenge is maintaining the efficiency of local training while allowing for global coordination.
Benjamin Holzschuh and Nils Thuerey addressed this in P3D by combining local convolutions with a "global context model." This hybrid approach allows for zero-shot scaling from small training patches to high-resolution 3D simulations (up to $512^3$) by linking separate patches through a coordination layer.
Similarly, Beomseok Kang and Minah Lee used Attentive Recurrent Neural Cellular Automata (AR-NCA) to learn local transition rules for systems like forest fires. By applying the same local rules to every cell, their model scaled to systems 64 times larger than the training set with zero performance loss.
Why It Matters
The shift toward zero-shot spatial scaling represents a move away from "brute force" deep learning. Instead of training on gigapixel images or massive 3D grids—which is often impossible due to hardware constraints—we can train on the fundamental building blocks of the problem.
As Theodor J. L. Forgaard notes in the context of Earth Observation (THOR), this allows us to use weights optimized on small, localized crops to process massive geographic windows. By prioritizing local equivariance, we unlock the ability to solve global-scale problems using only small-scale data.